experiments/baseline_lm_qa.py

import os
import json
import gc

import torch
import pandas as pd
from datasets import load_dataset
from lmppl import EncoderDecoderLM, LM, OpenAI

OPENAI_API_KEY = os.getenv("OPENAI_API_KEY", None)

prompt_dict = {
    "is friend/ally of": "entities that are friends or allies",
    "is competitor/rival of": "entities that are competitors or rivals",
    "is known for": "examples of what entities are known for",
    "is influenced by": "what has influenced different entities",
    "is similar to": "examples of entities that are similar"
}
data = load_dataset("cardiffnlp/relentless", split="test")
full_result = []
for lm, ppl_class, batch, pretty_name in [
    ("google/flan-ul2", EncoderDecoderLM, 1, "Flan-UL2"),
    ("google/flan-t5-xxl", EncoderDecoderLM, 1, "Flan-T5\textsubscript{XXL}"),
    ("google/flan-t5-xl", EncoderDecoderLM, 1, "Flan-T5\textsubscript{XL}"),
    ("google/flan-t5-large", EncoderDecoderLM, 32, "Flan-T5\textsubscript{LARGE}"),
    ("google/flan-t5-base", EncoderDecoderLM, 128, "Flan-T5\textsubscript{BASE}"),
    ("google/flan-t5-small", EncoderDecoderLM, 256, "Flan-T5\textsubscript{SMALL}"),
    ("t5-11b", EncoderDecoderLM, 1, "T5\textsubscript{XXL}"),
    ("t5-3b", EncoderDecoderLM, 1, "T5\textsubscript{XL}"),
    ("t5-large", EncoderDecoderLM, 32, "T5\textsubscript{LARGE}"),
    ("t5-base", EncoderDecoderLM, 128, "T5\textsubscript{BASE}"),
    ("t5-small", EncoderDecoderLM, 256, "T5\textsubscript{SMALL}"),
    # ("facebook/opt-66b", LM, 1, "OPT\textsubscript{66B}"),
    ("facebook/opt-30b", LM, 1, "OPT\textsubscript{30B}"),
    ("facebook/opt-13b", LM, 1, "OPT\textsubscript{13B}"),
    ("facebook/opt-6.7b", LM, 1, "OPT\textsubscript{6.7B}"),
    ("facebook/opt-2.7b", LM, 1, "OPT\textsubscript{2.7B}"),
    ("facebook/opt-1.3b", LM, 1, "OPT\textsubscript{1.3B}"),
    ("facebook/opt-350m", LM, 128, "OPT\textsubscript{350M}"),
    ("facebook/opt-125m", LM, 256, "OPT\textsubscript{125M}"),
    ("facebook/opt-iml-30b", LM, 1, "OPT-IML\textsubscript{30B}"),
    ("facebook/opt-iml-1.3b", LM, 1, "OPT-IML\textsubscript{1.3B}"),
    ("facebook/opt-iml-max-30b", LM, 1, "OPT-IML\textsubscript{MAX-30B}"),
    ("facebook/opt-iml-max-1.3b", LM, 1, "OPT-IML\textsubscript{MAX-1.3B}"),
    # ("davinci", OpenAI, None, "GPT-3\textsubscript{davinci}")
]:
    os.makedirs(f"experiments/results/lm_qa/{os.path.basename(lm)}", exist_ok=True)
    scorer = None
    for d in data:
        ppl_file = f"experiments/results/lm_qa/{os.path.basename(lm)}/ppl.{d['relation_type'].replace(' ', '_').replace('/', '__')}.jsonl"

        if not os.path.exists(ppl_file):

            if scorer is None:
                if ppl_class is OpenAI:
                    scorer = ppl_class(OPENAI_API_KEY, model=lm)
                else:
                    scorer = ppl_class(lm, device_map='auto', low_cpu_mem_usage=True, offload_folder=f"./offload_folder/{os.path.basename(lm)}")

            proto = ",".join([f'["{a}", "{b}"]' for a, b in d['prototypical_examples']])
            prefix = f"Answer the question by yes or no. We know that {proto} are examples of {prompt_dict[d['relation_type']]}."
            if ppl_class is LM or ppl_class is OpenAI:
                prompt_input = [f'{prefix} Are ["{x}", "{y}"] {prompt_dict[d["relation_type"]]} as well?\n yes' for x, y in d['pairs']]
                ppl = scorer.get_perplexity(input_texts=prompt_input, batch=batch)
                output = [{"perplexity": p, "input": i, "output": ""} for p, i in zip(ppl, prompt_input)]
            elif ppl_class is EncoderDecoderLM:
                prompt_input = [f'{prefix} Are ["{x}", "{y}"] {prompt_dict[d["relation_type"]]} as well?' for x, y in d['pairs']]
                ppl = scorer.get_perplexity(input_texts=prompt_input, output_texts=["yes"] * len(prompt_input), batch=batch)
                output = [{"perplexity": p, "input": o, "output": "yes"} for p, o in zip(ppl, prompt_input)]
            else:
                raise ValueError(f"Unknown class {ppl_class}")

            with open(ppl_file, "w") as f:
                f.write("\n".join([json.dumps(i) for i in output]))

        with open(ppl_file) as f:
            ppl = [json.loads(i)['perplexity'] for i in f.read().split("\n") if len(i) > 0]
        true_rank = d['ranks']
        assert len(true_rank) == len(ppl), f"Mismatch in number of examples: {len(true_rank)} vs {len(ppl)}"
        rank_map = {p: n for n, p in enumerate(sorted(ppl), 1)}
        prediction = [rank_map[p] for p in ppl]
        tmp = pd.DataFrame([true_rank, prediction], index=['true', 'pred']).T
        cor = tmp.corr("spearman").values[0, 1]
        full_result.append({"model": pretty_name, "relation_type": d['relation_type'], "correlation": cor})
    del scorer
    gc.collect()
    torch.cuda.empty_cache()

df = pd.DataFrame(full_result)
models = df['model'].unique()
df = df.pivot(columns="relation_type", index="model", values="correlation")
df = df.T[models].T
df['average'] = df.mean(1)
df.to_csv("results/lm_qa/lm.csv")
df = (100 * df).round()
print(df.to_markdown())
print(df.to_latex(escape=False))